fix: replace blocking waitUntilExit with async terminationHandler to prevent thread starvation

[irena327]

Summary

Replace process.waitUntilExit() with process.terminationHandler + CheckedContinuation to avoid blocking the cooperative thread pool

Problem

CommandRunner.run() executes inside a Task.detached, which runs on Swift's cooperative thread pool. The call to process.waitUntilExit() blocks that thread until the subprocess finishes. When multiple commands run concurrently (e.g., during manifest loading in a large monorepo), each blocked thread reduces pool capacity. Once all pool threads are occupied by waitUntilExit() calls, the stdoutTask and stderrTask pipe-reading tasks can never be scheduled. The subprocesses then fill their pipe buffers and block on write, while waitUntilExit() waits for processes that can never finish — a classic thread starvation deadlock.

This causes tuist generate and tuist test to hang indefinitely at "Loading and constructing the graph" in large repos.

Fix

Replace the blocking waitUntilExit() with an async-friendly terminationHandler that suspends the task via CheckedContinuation. This frees the pool thread while waiting for the process to exit, allowing pipe-draining tasks to be scheduled normally.

[fortmarek]

Added a regression test in CommandRunnerRaceTests that exercises the pipe-buffer starvation case described in the PR.

What changed:

• Added runsManyConcurrentCommandsWithLargeOutput_successfully.
• The test starts 64 concurrent /bin/sh commands.
• Each command writes 256 KiB to stdout and 256 KiB to stderr.
• The test asserts that CommandRunner drains the full byte count from both streams.

Local verification:

• Temporarily reverted the implementation back to process.run(); process.waitUntilExit() while keeping the new test. The build completed, the test started, and then it hung, reproducing the deadlock shape this PR is fixing.
• Restored the async terminationHandler + continuation implementation from this PR and reran the same test successfully:
  • swift test --scratch-path /private/tmp/command-pr-249-regression-build --filter CommandRunnerRaceTests.runsManyConcurrentCommandsWithLargeOutput_successfully
  • Passed in 0.212s on the final run.

This gives us a TDD-style regression check: old behavior hangs, fixed behavior drains the pipes and completes.